By Matt Day | Updated on Oct 12, 2026 at 02:00 PM
Microsoft Corp. is gearing up to install the latest version of its own chips in data centers, with the aim of powering its AI software using mostly homegrown hardware.
The company expects to stock multiple gigawatts of its Maia chips in its server farms, with the newest iteration, the Maia 300, starting to roll out next year, Microsoft said on Monday. The semiconductors are designed to handle inference — using an AI model for tasks such as summarizing a conference call or producing answers to a research question.
Microsoft, Amazon.com Inc. and Alphabet Inc.’s Google are all creating in-house artificial intelligence accelerators, an effort that could potentially let them reduce their reliance on Nvidia Corp. hardware. Amazon and Google began their chip-making efforts earlier than Microsoft and are perceived to hold a lead over the world’s largest software maker.
Microsoft is betting that by tailoring chips to its own data centers, it can run inference more efficiently and avoid some of the expense of stocking Nvidia’s market-leading graphics processing units, which power most of that work at Microsoft today.
“We are confident we have built the best product for these workloads,” said Rani Borkar, the Microsoft executive who oversees data center hardware, including the chip design teams. “We are backing that confidence by deploying Maia at gigawatt scale starting next year.”
Microsoft began installing an earlier version called Maia 200 in its data centers early this year. That chip is currently helping to power the company’s Copilot AI software, among other tasks, Borkar said.
Maia 200 also runs OpenAI and Microsoft AI team models, along with open-source model families. It’s in data centers in the US, Europe and Australia, where it powers AI capabilities in Microsoft Teams and Copilot, among other products.
Microsoft expects the more powerful Maia 300 to deliver 40% more tokens per megawatt than leading GPU-based systems. Tokens, which in practice tend to represent a word or a piece of one, are the granular unit of AI output.
Borkar said the company isn’t currently planning to rent the underlying computing power of Maia chips directly to its cloud customers, a step that Google and Amazon have taken with their own AI chips.
“Our own AI services business is big enough that we want to serve ourselves first,” Borkar said.